varframe 1.2.1

Declarative DataFrame variable management with automatic DAG dependency resolution and ML model integration

VarFrame

Declarative DataFrame variable management with automatic DAG dependency resolution and ML model integration.

---

What is VarFrame?

VarFrame is a library that allows you to define DataFrame columns as Python classes rather than imperative scripts. It manages dependencies, types, and execution order automatically using a generic DAG (Directed Acyclic Graph) solver.

It is designed for complex, production-grade data pipelines where traceability, correctness, and structure are more important than raw implementation speed.

graph TD
    Raw[Raw DataFrame] -->|extracts| Base[BaseVariable]
    Base -->|inputs| Derived[DerivedVariable]
    Base -->|features| Model[ML Model]
    Derived -->|features| Model
    Model -->|predicts| Pred[ModelVariable]
    Pred -->|inputs| Ensemble[Ensemble Model]
    Ensemble -->|predicts| Final[Final Prediction]

    style Raw fill:#e1f5fe,stroke:#01579b,stroke-width:2px
    style Base fill:#fff9c4,stroke:#fbc02d,stroke-width:2px
    style Derived fill:#e0f2f1,stroke:#00695c,stroke-width:2px
    style Model fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px
    style Pred fill:#fce4ec,stroke:#c2185b,stroke-width:2px
    style Ensemble fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px

Why VarFrame?

The Problem with Traditional Scripts

In traditional pandas scripts (df['b'] = df['a'] + 1), logic is often:

• Fragile: Reordering cells or lines breaks dependencies silently.
• Opaque: It's hard to tell exactly which columns are needed effectively.
• Hard to Test: You have to test "intermediate states" of a large dataframe.

The VarFrame Solution

VarFrame treats variables as definitions (Classes) rather than steps.

Feature	VarFrame	Traditional Script
Dependency Resolution	Automatic (DAG). Order doesn't matter; the framework solves it.	Manual. You must order operations correctly yourself.
Logic Encapsulation	Logic, metadata, and types live in one Class. Self-documenting.	Distributed across scripts. logic often mixed with execution.
ML Integration	Models are just "Computed Variables". Predictions are treated like any other column.	Often separate "training" and "inference" pipelines.
Testing	Unit test single calculate(df) methods in isolation.	Integration testing entire scripts is required.

Best For

• Feature Stores: Reuse definitions across training and serving.
• Complex DAGs: When variable F depends on E, which depends on D, C, and B...
• Ensemble/Stacking: Where model predictions feed into other models (see examples/ensemble_demo.py).

Installation

pip install varframe           # Core only (pandas)
pip install varframe[ml]       # + scikit-learn, joblib
pip install varframe[all]      # Everything

Quick Start

1. Define Variables

from varframe import BaseVariable, DerivedVariable, VarFrame

# Map a raw column with type enforcement
class Lap(BaseVariable):
    """Current lap number."""
    name = "lap"
    raw_column = "lap_num"
    dtype = "int"

class Gap(BaseVariable):
    """Gap to leader in seconds."""
    name = "gap"
    raw_column = "gap_to_leader"
    dtype = "float"

# Create a computed column with dependencies
class GapDelta(DerivedVariable):
    """Change in gap from previous row."""
    name = "gap_delta"
    dependencies = [Gap]
    
    @classmethod
    def calculate(cls, df):
        return df["gap"] - df["gap"].shift(1)

2. Create a VarFrame

import pandas as pd

# Raw data with original column names
df_raw = pd.DataFrame({
    "lap_num": [1, 2, 3],
    "gap_to_leader": [0.0, 1.2, 0.8]
})

# Create VarFrame - columns are computed automatically
# Dependencies are resolvd automatically!
vf = VarFrame(df_raw, [Lap, Gap, GapDelta])

print(vf)
#    lap  gap  gap_delta
# 0    1  0.0        NaN
# 1    2  1.2        1.2
# 2    3  0.8       -0.4

3. Access Variables

# By name
vf["gap"]

# By class
vf[Gap]

# Multiple variables
vf[[Lap, Gap]]

# Filter by type
vf.filter_by_type(DerivedVariable)  # Only computed columns

ML Model Integration

Define models declaratively and use predictions as variables:

from varframe import BaseModel, ModelVariable
from sklearn.ensemble import RandomForestRegressor

class GapPredictor(BaseModel):
    """Predicts future gap based on features."""
    name = "gap_predictor"
    inputs = [Lap, Gap]
    target = GapDelta
    model_class = RandomForestRegressor
    hyperparameters = {"n_estimators": 100, "max_depth": 5}

# Train the model
GapPredictor.train(training_vf)

# Use predictions as a variable
class PredictedGapDelta(ModelVariable):
    name = "predicted_gap_delta"
    model_class = GapPredictor

vf.add_variables(PredictedGapDelta)

Optimization & Export

Lazy Loading

Optimize memory by marking variables as lazy = True. They are computed on-demand and not stored in the DataFrame.

class HugeFeature(DerivedVariable):
    lazy = True
    dependencies = [RawData]

    @classmethod
    def calculate(cls, df):
        return df["raw"] * 1000

Flexible Views

Export specific subsets of data using vf.view():

# Export only base variables
df_base = vf.view(include=["base"])

# Export specific variables (computes lazy vars on demand)
df_custom = vf.view(variables=[HugeFeature])

API Reference

Variable Classes

Class	Purpose
BaseVariable	Maps a raw column (with optional dtype conversion)
DerivedVariable	Computed from other variables. Set lazy=True for on-demand computation.
ModelVariable	Predictions from an ML model

VarFrame Methods

Method	Description
add_variables(*vars, compute=True)	Compute and add new variables (or register if compute=False)
add_variable(*vars)	Alias for add_variables(*vars)
filter_by_type(type)	Filter to BaseVariable or DerivedVariable only
get_variable(name)	Get variable class by name
view(include=..., variables=...)	Export DataFrame with specific variables (handles lazy computation)
list_variables()	List all variable names
describe_variables()	Summary DataFrame of all variables
to_pandas() / to_ml()	Convert to plain DataFrame for ML pipelines

BaseModel Methods

Method	Description
train(vf)	Train on a VarFrame
predict(vf)	Generate predictions
evaluate(vf)	Compute metrics
save(path) / load(path)	Persist and restore model

License

MIT